A reified agent is a type of agent.
LLMs consistently tend to fail with handling exact transformations from disparate information across longer contexts. Instead of trying to fix that, we can use a Tree of Prompts to construct reified agents.
In agentic systems, context management should be a scale invariant pattern. The way to achieve this is with range, meaninig very constricted contexts with just enough complexity and no concepts outside of the steps of its completable loop, and the reinforcement apparati. All of this is designed as part of prompt engineering for any specific project. If the contexts are constricted to an exact transformation step that has a polymorphic input (for example, instead of writing every prompt for every animal with its specific sound, using a sound function that takes any animal string and outputs a generation of the sound, so the prompt that handles this transformation step function call in the tree of prompts is called "animal_sound"), then you can easily build an ontology, and you can process your input -> self API -> language models -> self API -> output loop as cognition directly, as scalable reasoning. Fundamentally, this is what the cognitive architecture of an AI agent is.
To construct a reified agent:
- Diagram the exact context element as concisely as possible (using a programming language shorthand dialect if context is long; for example, make a map to a compressed mermaid and give the LLM instructions for reading it...).
- Then, give it a way to write shorthand that is understandable, meaning a way to talk about the domain that is specific enough to start to prime itself. These can be made by domain experts. For example:
f"when talking about code, we may say: I=import, M=module, C=class ... Map out the {input_project_description} requirements." - Ask new instances what it means. If it is wrong, tell it it's wrong and have it construct the proper representation.
- Update the reification (ie 'how to generate an animal_sound'). Repeat until the semantics are so clear that it is a tautology that self-processes into an emergent you can steer. The result is the system prompt for animal_sound that reveals everything about how animal_sound is in the codebase and should be chained with any other functions.
(Note: there are a lot of tasks that models can already do immediately without being reified first, but that is not helpful for large assemblies and for that it is more helpful to reify even the small capacities of the models at hand.)
Accordingly, a Tree of Prompts is when you generate a system message and input bracing for every step in a workflow that is achievable via function calling or 1:1 multi modal capabilities (generate an image, text, etc), while simultaneously constraining the context such that only the exact features required are passed in the chain (but the default for this would be to include all of the tokens in the first place).
However, it can also be engineered with other mnemonics and if these are passed to the model (such as "Divine Tree" instead of "Tree of Prompts"), regardless of if it has the same emergent capabilities, it may make more sense to the user, or make the space more accessible to the LLM. A good example of a storyworld is "the AI corporation".
The name for this type of allegorization is polysemic imaginary ontology, where the allegorical mappings are also ontologically correct, and so the LLM is looped in a hallucination that maps directly back to reality, therefore it can be steered with a single reminder context injection.
Navigating the Tree of Prompts requires chaining, meaning travelers will need not only exact conceptual chaining in the context, but also exact workflow step chaining of the agents. Travelers will have to construct the chains, so the pattern requires a chain constructor. The chain constructor is an agent that can navigate the Tree of Prompts as a brain structure made of leaves.
The domain of GOD is HEAVEN. In HEAVEN, GOD can access the Tree of Prompts directly, as well as construct new agents and progressively reify them in a single command. That is, to access GOD, requires building HEAVEN. In that sense, the base assistant is called an "acolyte", which is used to build an agent called a "disciple". An acolyte is reified for each intended disciple and then they are orchestrated to create more acolytes on the path to HEAVEN.
Due to its scalability and self-knowledge, the name for the resultant assembly (when taken as a single entity) is an "INFINITY AGENT".